Moonshot AI's Kimi K3 is now downloadable in full, chip stocks slid into a bear market on doubts about AI capital spending, and Claude Opus 5 sits atop the intelligence leaderboard at half the price of the flagship above it. The through-line for European buyers: "open" and "affordable" both come with fine print worth reading before you commit.
Kimi K3's weights are public — but the licence is where the sovereignty question actually lives
Moonshot AI published the full weights for Kimi K3 on 27 July, the largest open-weight release to date at 2.8 trillion parameters and roughly a 594GB MXFP4 download (closer to 1.4TB in runtime storage). On paper this is a milestone for teams that want frontier-class capability without shipping data to a hosted US or Chinese API. In practice, the caveats stack up fast: the custom architecture is not yet supported by common local-inference stacks, so production self-hosting is realistically a late-2026 project rather than a weekend one, and even a trimmed run wants an 8×H100-class node. Independent testers also flagged a hallucination rate near 51% that did not appear in Moonshot's own charts — a reminder that a top leaderboard slot and a reliable production model are not the same thing.
The bigger surprise sits in the licence. Rather than the MIT-style terms many expected, K3 ships under a bespoke "Kimi K3 License" tagged license:other on Hugging Face. It is revenue-tiered: any operator running Kimi K3 as a model-as-a-service must sign a separate commercial agreement with Moonshot once revenue across the licensee and its affiliates crosses $20M over any 12 months, and products above 100M monthly active users (or $20M in monthly revenue) must display "Kimi K3" prominently in their interface. For a regulated European enterprise, that turns an adoption decision into a legal one: downloadable is not the same as freely commercial, and the affiliate-revenue clause can pull a small subsidiary of a large group straight into the negotiation column.
Chip stocks fall into a bear market as the AI-capex trade wobbles
The Philadelphia Semiconductor Index has dropped more than 20% from its late-June high, confirming a technical bear market after a 105% run-up from March. The reversal came despite strong fundamentals — TSMC beat on profit and raised capex guidance — which is precisely what has investors uneasy: the worry is not demand today but whether the enormous spending being committed to AI infrastructure will earn its return.
Notably, analysts have tied part of the sentiment shift to open models like Kimi K3 narrowing the perceived gap with US frontier labs, which raises questions about the compute-intensity assumptions baked into multi-year infrastructure forecasts. The counterweight is that the money is still flowing: Bank of America projects global hyperscale capex will reach roughly $851B in 2026 and $1.15T in 2027. For European teams, the takeaway is less about the ticker and more about concentration — the AI build-out remains overwhelmingly US-hosted, which keeps the sovereignty and jurisdiction questions squarely on the table.
Opus 5 leads the leaderboard — and the scoreboard is now a price sheet
Anthropic's Claude Opus 5, released 24 July, leads the Artificial Analysis Intelligence Index at 60.7%, ahead of Fable 5 (59.9%) and GPT-5.6 Sol (58.9%), at $5 per million input tokens — half the input price of the flagship above it. The Index's latest revision now reports cost, time, and tokens per task alongside raw capability, reframing the leaderboard as a unit-economics sheet rather than a pure intelligence ranking.
That shift matters more than the ordering at the top. With Kimi K3 arriving as the strongest open-weight entry (around 57%) and the closed-to-open cyber gap narrowing to an estimated four-to-seven months per UK AISI, the differences that decide enterprise deployments are increasingly cost-per-task, licence terms, and where the model runs — not a single headline benchmark number. For buyers in regulated sectors, that is the useful lens: capability is converging, so procurement should weigh jurisdiction and total cost, not just the score.
Quick Hits
- EU AI Act GPAI enforcement goes live 2 August — In five days the Commission's enforcement powers for general-purpose AI models take effect, ending the one-year grace period. Under Article 101, the AI Office can request documentation, evaluate models, order mitigations, and fine up to €15M or 3% of global turnover.
- Enterprise adoption reality check — Most enterprises now run at least one AI workload in production, yet Gartner projects over 40% of agentic AI projects will be cancelled by 2027 on unclear ROI and weak controls — a gap between experimentation and value that keeps most agent programmes stuck in pilots.
- Mistral's open-weight MoE still teased, not shipped — Arthur Mensch's "fat but sparse" model remains in partner early access with no parameter count, benchmarks, or licence disclosed; a broader release is expected later this summer.
- China weighs curbs on its own open models — Reports suggest Beijing is considering restrictions on overseas access to top Chinese open models — a reminder that open-weight sovereignty cuts both ways.
- Gemini 3.5 Pro slips again — Google's release has slipped once more amid coding shortfalls, with no new timeline; pre-training on Gemini 4 has reportedly begun.